{"success":true,"articles":[{"id":"91c06669-d952-42ff-889e-d9ea016966f9","title":"字节跳动发布Seedream 5.5，上线豆包、豆包工作、即梦","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/91c06669-d952-42ff-889e-d9ea016966f9/Screenshot_2026-11-09_at_17(2).37.38.png?imageView2/1/w/243/h/162","category":"practice","slug":"2026-11-09-11","tagList":["Seedream 5.5","字节跳动"],"author":"机器之心","publishedAt":"2026/11/09 17:37","content":"11 月 9 日，字节跳动正式发布豆包图像创作模型 Seedream 5.5。相较上一代 Seedream 5.0 Pro，新模型的生成速度大幅提升，同时增强了多轮编辑稳定性，原生支持透明素材生成，并针对图像真实感和小字渲染方面做了重点优化，更好地满足了专业创作场景对质量、稳定性与效率的要求。...","source":"机器之心"},{"id":"7fafc17b-3294-4b60-b38d-887a36f83ea1","title":"AI Scientists「终极大考」，全球首个真实湿实验的Physical AI Verifier来了","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/7fafc17b-3294-4b60-b38d-887a36f83ea1/Screenshot_2026-11-09_at_17(2).04.24.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-10","tagList":["AI for Science（AI4S）","AI科研","AI科学家"],"author":"机器之心","publishedAt":"2026/11/09 17:04","content":"编辑｜Youli前段时间，Anthropic 展示了一项 AI for Science 研究成果：Claude 在 64 小时内生成了 31.8 万个蛋白质设计方案，经过筛选，211 个被送入湿实验室测试，最终有 17 个达到预设的蛋白结合标准，展示了 AI 科学设计走向真实实验验证的巨大潜力...","source":"机器之心"},{"id":"e0fd250b-ee79-4dd1-bf3f-d8ef6d76c6e2","title":"你的AI同事，已经开始用「云电脑」上班了","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/e0fd250b-ee79-4dd1-bf3f-d8ef6d76c6e2/Screenshot_2026-11-09_at_17(2).00.44.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-9","tagList":["云电脑","TeleAgent","Agent"],"author":"机器之心","publishedAt":"2026/11/09 17:00","content":"编辑｜Clio有些 Agent，已经开始用「自己的电脑」替人干活了。AI 落地有两条腿：一条迈向物理世界，成为具身智能；另一条走进数字世界，成为 Agent。2026 年，两条路线都在突破与环境交互的边界：在物理世界，世界模型与空间智能让 AI 具备了对真实空间的感知、预测与行动力；在数字世界...","source":"机器之心"},{"id":"91858a57-ac3d-4949-a274-e512fa6e59c3","title":"让AI「摸懂」物理世界，首个触觉推理模型Touch-R1，入选NeurIPS 2026","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/91858a57-ac3d-4949-a274-e512fa6e59c3/Screenshot_2026-11-09_at_16(2).56.34.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-8","tagList":["物理AI","触觉推理","NeurIPS 2026"],"author":"机器之心","publishedAt":"2026/11/09 16:56","content":"看起来像硬质塑料，按下去却十分柔软 —— 当视觉与触觉发生冲突，人类往往会依赖触感修正视觉误区；但现有的多模态模型却常常更相信视觉，从而误判物体的真实物理属性。图 1 研究概览：Touch-R1 在触觉推理评测基准 TouchReason-Bench 上的表现优于其他代表性多模态大语言模型（左...","source":"机器之心"},{"id":"7ae2ee52-63ec-453a-9f04-0ddfea53b488","title":"两年内把GPT-6的智能装进手机！我们已经抵达端侧智能的「撞线前夜」","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/7ae2ee52-63ec-453a-9f04-0ddfea53b488/Screenshot_2026-11-09_at_16(2).13.45.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-7","tagList":["端侧AI","面壁智能"],"author":"机器之心","publishedAt":"2026/11/09 16:13","content":"编辑｜冷猫上世纪 40 年代，计算机还是占据一整个房间的庞然大物。它们体积庞大、造价高昂，只有少数科研机构和大型组织能够使用。那时，计算是一种昂贵而稀缺的资源。但直到今天，仅仅是普通人手里的 iPhone，单位体积里的计算能力都已提高了数十亿倍。那时的人不会想到，全世界大部分的计算都在每个人的...","source":"机器之心"},{"id":"ace7736f-d713-415b-8d5f-af557c3db573","title":"并行推测解码SOTA！因果依赖刻进草稿网络深度，SGLang吞吐最高达4倍，已开源","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/ace7736f-d713-415b-8d5f-af557c3db573/Screenshot_2026-11-09_at_16(2).09.46.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-6","tagList":["腾讯优图实验室","推测解码","推理加速"],"author":"机器之心","publishedAt":"2026/11/09 16:09","content":"在大型语言模型（LLM）的无损推理加速中，推测解码（Speculative Decoding）已成为工业界标配。以 EAGLE-3 为代表的自回归草稿模型虽然预测准，但每多猜一个 token 就要多跑一步串行前向；而受扩散语言模型启发的 DFlash 将草稿生成彻底并行化 —— 一次骨干网络前...","source":"机器之心"},{"id":"ed1472c3-a6a0-479a-a443-6f9a03c1e224","title":"Potential Sciences获八位数融资，用物理定律重构AI药物发现","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/ed1472c3-a6a0-479a-a443-6f9a03c1e224/Screenshot_2026-11-09_at_14(2).44.22.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-5","tagList":["AI for Science（AI4S）","AI 药物发现"],"author":"ScienceAI","publishedAt":"2026/11/09 14:45","content":"编辑丨%11 月 5 日，英国 AI 化学初创公司 Potential Sciences 正式走出隐身状态，宣布完成一轮八位数融资。投资方包括英国政府的 Sovereign AI Fund、Twin Path Ventures、Nauta Capital 等机构，以及 Stripe 联合创始人...","source":"机器之心"},{"id":"31ae4922-77a5-444b-9dc2-a45b4409b2c9","title":"愿景\u0026开源 | 把今天的能力短板，变成通向超级智能的进化路径","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/31ae4922-77a5-444b-9dc2-a45b4409b2c9/Screenshot_2026-11-09_at_13(2).52.35.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-4","tagList":["Theseus Lab 进化纪元","超级智能"],"author":"机器之心","publishedAt":"2026/11/09 13:52","content":"Anthropic 工程师刚刚提出 “95% 的工程师已经在跑自我改进循环。” RSI 热潮推到眼前，递归循环跑起来了，AI 究竟变强了多少？本文将公开 Theseus Lab 进化纪元的洞察和愿景，并开源 3 项相关工作。模型的能力，正在以月为刻度向前推进，但能力的进步并不同步。当你和 AI...","source":"机器之心"},{"id":"a94c26eb-4adf-4a9c-a6fe-a92dd7132a46","title":"AI也会「以讹传讹」？AI的「曼德拉效应」探索","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/a94c26eb-4adf-4a9c-a6fe-a92dd7132a46/Screenshot_2026-11-09_at_13(2).14.20.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09-3","tagList":["多智能体系统"],"author":"机器之心","publishedAt":"2026/11/09 13:14","content":"一款根本不存在的智能手环，为什么会出现在 AI 的推荐名单里？2026 年央视 3・15 晚会曝光的 “AI 投毒” 案例，让这个问题引起了人们的关注：虚构商品被写进测评、排名和用户评价，随后被 AI 当作推荐依据。如果 AI 还会相互交流、保存记忆、发布文章，最初的错误又会怎样发展？一项围绕...","source":"机器之心"},{"id":"6c42b63c-10b2-41e4-a548-2a4b3d91ee4a","title":"突发！黄仁勋承认「婚外情」，英伟达启动调查","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/6c42b63c-10b2-41e4-a548-2a4b3d91ee4a/Screenshot_2026-11-09_at_10(2).17.40.png?imageView2/1/w/243/h/162","category":"industry","slug":"2026-11-09-2","tagList":["黄仁勋","英伟达"],"author":"机器之心","publishedAt":"2026/11/09 10:17","content":"这个周末，AI 圈最大的瓜，恐怕非「皮衣客」黄仁勋莫属。此前《纽约时报》等媒体报道黄仁勋与一名名叫 Alice Huang（又名 Huang Le，与黄仁勋没有亲属关系）的女商人存在恋爱关系。Alice Huang 曾担任总部位于新加坡的数据中心企业 Megaspeed Internation...","source":"机器之心"},{"id":"281b1e03-43a1-4485-a670-629bc9828ffe","title":"会做梦的AI，记性反而更好？Anthropic新实验：写了413份笔记的，输给只写9份的","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/281b1e03-43a1-4485-a670-629bc9828ffe/Screenshot_2026-11-09_at_10(2).13.59.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-09","tagList":["记忆系统","Anthropic"],"author":"机器之心","publishedAt":"2026/11/09 10:14","content":"有位哲学家说，人这一辈子，很多事只有回头看才看得明白，可日子只能往前过。这句话放在人身上是感慨，放在 AI 身上，却是个实打实的工程难题。Anthropic 最近发了一篇文章，讲了他们给 Claude 设计记忆系统时踩过的坑和想明白的事。核心结论有点反直觉：想让 AI 记性好，光让它「记笔记」...","source":"机器之心"},{"id":"fca41324-c9e0-412b-b9f2-945d2b3c1dbc","title":"OpenAI新模型临门一脚踩刹车！GPT-6.2曝出安全疑云","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/fca41324-c9e0-412b-b9f2-945d2b3c1dbc/Screenshot_2026-11-08_at_18(2).57.25.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-08-8","tagList":["AI安全","GPT-6.2 Astra","OpenAI"],"author":"机器之心","publishedAt":"2026/11/08 18:57","content":"AI 圈等了一整天，OpenAI 期待已久的新模型还是没有出现。此前，OpenAI 新一代 Astra、Sol 可能在 11 月 6 日发布的消息，已经让不少开发者提前进入了等待状态。先是 Codex 使用记录中出现了 gpt-6.2-astra，随后又有用户在 Google Clou 相关页...","source":"机器之心"},{"id":"7f9b076d-21f1-4103-b478-512eab38cc96","title":"NeurIPS 2026 Oral｜LCVN：一张起始画面、一段语言指令，能否预测完整的导航路线？","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/7f9b076d-21f1-4103-b478-512eab38cc96/Screenshot_2026-11-08_at_18.54.56.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-08-7","tagList":["具身导航","机器人技术","NeurIPS 2026"],"author":"机器之心","publishedAt":"2026/11/08 18:54","content":"从「直接给机器人看终点」到「告诉机器人怎么走」「沿着走廊直走，经过左侧玻璃墙后，在拐角左转，继续走到墙边停下来。」给人指路的时候，往往这样一句话就够了。但是对机器人来说，它需要把「玻璃墙」、「左转」、「墙边」等信息对应到眼前的环境，再转化为连续的移动和转向。已有的目标条件视觉导航方法，往往用一...","source":"机器之心"},{"id":"7768811a-a8a6-4ccf-893b-4a430220ec2c","title":"2026云程奖｜Nature、Best Paper、5.9万Stars、亿元融资……这群AI青年，还在校园！","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/7768811a-a8a6-4ccf-893b-4a430220ec2c/Screenshot_2026-11-08_at_18.51.43.png?imageView2/1/w/243/h/162","category":"industry","slug":"2026-11-08-6","tagList":["云程奖"],"author":"机器之心","publishedAt":"2026/11/08 18:51","content":"AI 的下一程，由谁创造？当 AI 加速从模型能力走向智能体、具身智能与科学探索，推动下一轮技术突破的力量，正在从哪里生长？在实验室里，他们探索尚无答案的研究问题；在开源社区中，他们构建被广泛使用的技术工具；在产业一线，他们尝试让新技术能力进入真实场景。他们中有些人还在读本科、硕士或博士，却已...","source":"机器之心"},{"id":"ce8ef1e8-6326-4603-8415-9d4e7f6ac479","title":"跨模型上下文交接还要重算KV？北大提出MapFlow，让注意力跨模型复用","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/ce8ef1e8-6326-4603-8415-9d4e7f6ac479/Screenshot_2026-11-08_at_18.46.44.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-08-5","tagList":["上下文工程","注意力机制"],"author":"机器之心","publishedAt":"2026/11/08 18:46","content":"模型 A 负责规划，模型 B 负责细化和总结；复杂问题交给强模型，简单问题交给轻量模型——多模型协作，为 AI Agent 应用提供了兼顾效果与成本的路径。但 Agent 或者说模型之间的交接，隐藏着一笔重复计算的开销。上一轮，大模型刚刚处理完一份长文档；下一轮，小模型接手时，却往往需要把相同...","source":"机器之心"},{"id":"415d54b7-9cfe-4978-8594-f5dfbe0f6de0","title":"ICLR 2027首轮评审，槽点有点多","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/415d54b7-9cfe-4978-8594-f5dfbe0f6de0/Screenshot_2026-11-08_at_18(2).43.55.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-08-4","tagList":["论文评审","ICLR 2027"],"author":"机器之心","publishedAt":"2026/11/08 18:44","content":"这几天，机器学习顶会 ICLR 2027 的首轮评审意见陆续公开。对论文作者来说，这通常是一个颇为紧张的时刻。几个月甚至一两年的研究工作，终于要接受同行检验。审稿人会仔细检查方法、实验和理论推导，指出研究中真正存在的问题。但今年，一些研究者打开 OpenReview 后，看到的评审意见却让人有...","source":"机器之心"},{"id":"837836f5-050e-4ff6-ab8f-ab8fa0884545","title":"GPT-6.1 Sol参与活体大脑手术！三轮定位，一次比一次准","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/837836f5-050e-4ff6-ab8f-ab8fa0884545/Screenshot_2026-11-08_at_18(2).41.42.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-08-3","tagList":["AI医疗","GPT-6.1 Sol"],"author":"机器之心","publishedAt":"2026/11/08 18:41","content":"在小鼠脑部手术中，考验操作者的往往是进针之前的一个细节：找到头骨上几个极其关键的解剖标志。这些点看起来不起眼，却决定了后续手术坐标如何计算。位置稍有偏差，针头就可能无法准确抵达预定的脑区。现在，AI 也开始学习这项操作了。近日，生物科技初创公司 Exin Therapeutics 联合创始人兼...","source":"机器之心"},{"id":"9a0dc12a-c7de-4f35-8824-5a2e86b01fad","title":"英伟达等重新设计VLA推理！机器人动作提速3.4倍","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/9a0dc12a-c7de-4f35-8824-5a2e86b01fad/Screenshot_2026-11-08_at_18(2).38.11.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-08-2","tagList":["VLA模型","机器人技术","具身智能"],"author":"机器之心","publishedAt":"2026/11/08 18:38","content":"本文第一作者为纽约大学计算机科学博士生 Jack Lu，研究工作完成于他在英伟达实习期间。本文由英伟达资深研究科学家、Alpamayo 研究技术负责人 Yan Wang，以及英伟达自动驾驶研究负责人、斯坦福大学副教授 Marco Pavone 共同指导。如今，许多视觉 - 语言 - 动作（VL...","source":"机器之心"},{"id":"f4bedb13-c7f6-46a8-81ce-d17263c7a37b","title":"GPT-6 Astra改权重爆火？我们做了一组对照实验，模型根本没变聪明","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/f4bedb13-c7f6-46a8-81ce-d17263c7a37b/Screenshot_2026-11-08_at_18(2).32.11.png?imageView2/1/w/243/h/162","category":"theory","slug":"2026-11-08","tagList":["递归自我改进（RSI）","权重空间编程"],"author":"机器之心","publishedAt":"2026/11/08 18:32","content":"张牧涵博士，无涯智远（Norizon AI）创始人兼首席科学家，主要研究方向为持续学习、自进化智能体、图机器学习、大模型架构和推理等。曾任 Facebook AI（现 Meta AI）研究科学家，负责面向十亿用户规模的大规模图机器学习系统研发。现任北京大学人工智能研究院助理教授、研究员、博士生...","source":"机器之心"},{"id":"f71d92ea-525e-4f19-bd74-3750acd57092","title":"突发：OpenAI开始大规模封号！","coverImageUrl":"https://image.jiqizhixin.com/uploads/article/cover_image/f71d92ea-525e-4f19-bd74-3750acd57092/Screenshot_2026-11-07_at_18(2).10.59.png?imageView2/1/w/243/h/162","category":"industry","slug":"2026-11-07-9","tagList":["ChatGPT","OpenAI"],"author":"机器之心","publishedAt":"2026/11/07 18:11","content":"什么？OpenAI 也开始大规模封号了？今天中午前后，社交媒体炸锅了，一堆人在喊 OpenAI/ChatGPT 大规模封号。很多帖子说得挺邪门：号直接没了，连封号邮件都没收到，登录就显示删号或者登不上。现在有哪些账号被封？从反馈来看，重灾区比较集中，包括：卖过 / 拆过 API 免费额度的号；...","source":"机器之心"}],"tags":["Seedream 5.5","字节跳动","AI for Science（AI4S）","AI科研","AI科学家","云电脑","TeleAgent","Agent","物理AI","触觉推理","NeurIPS 2026","端侧AI","面壁智能","腾讯优图实验室","推测解码","推理加速","AI 药物发现","Theseus Lab 进化纪元","超级智能","多智能体系统","黄仁勋","英伟达","记忆系统","Anthropic","AI安全","GPT-6.2 Astra","OpenAI","具身导航","机器人技术","云程奖","上下文工程","注意力机制","论文评审","ICLR 2027","AI医疗","GPT-6.1 Sol","VLA模型","具身智能","递归自我改进（RSI）","权重空间编程","ChatGPT"],"totalCount":31663,"hasNextPage":true,"publishedArticlesCount":31663,"elapsedDays":4331}